A Model Of Multiclass Pattern Recognition Based On Support Vector Machine

Classification theory of typical SVM is proposed for two class problem at first, now how to develop effectively to multi-class classification is an incomplete solving problem. At present, the approaches of solving multi-class classification problems has two, one is that the model is built once again based on typical SVM, another is that multi-class classification is break into two-class classification by strategies. The thesis improve the known multi-classification models by typical SVM and strategies, it has some research results as follows:①For excessive characteristic variables of classification pattern, characteristic variables are distilled and compressed by multiple linear discriminant analysis, then multi-class classification model based on SVM and multiple linear discriminant analysis is built up by classification decision-making rule of multi-class pattern recognition modes based on binary tree and one against one .The Model not only make better use of information both within and between categories, but also play a role of reducing dimension of pattern, reduce the calculating quantity, increase recognition capability. At last, the instances proved the model is feasibility and validity.②Considering of some training data of classifiable pattern must be difficult to get training sample, when to build the classified model, few pattern which lack it's training sample. To define the output as three different exports which are classifiable function, based on the classic pattern theory of support vector machines. To build up a dynamic model of multi-class classification by using the classifiable means which is the binary tree based on model of one against one. The model can not only solve rejected classification efficiently, but also recognize some new pattern sample and make use of the new sample information sufficiently, to amend the model of multi-class classification, to realize the procedure of dynamic classification. The feasibility and availability has been proved by the experiment of data.③Every sample has different contribution to beyond classifiable place, because of pollution of yawps in the reality. Especially in some cases, some samples can not be classified to certain classification definitely, which only attribute to one classification of some probability. So to research the classification of sample, not only to consider the two extreme things, one is the sample belong to one classification with probability 1, or the other is not belonging to one classification with probability 1. Enlightened by the decision of Bayesian, in this article, the author uses the posterior probability as the pattern labels of sample. Under the theory of the classification based on the posterior probability support vector machines, using the theory of the classification which is one against one, then, to build up a multi-class classification model based on posterior probability matrix by using the changing of posterior probability matrix .The model makes a method to solve the multi-class classification which is uncertainty. Contrasting with the formerly models of multi-class classification, its ability of classification is improved preferably. The feasibility and availability has been proved by the experiment of data.The thought of integration is used in studying the classification organs, a classifiable model based on measurement maybe is a good choice in the future.